diff options
Diffstat (limited to 'src/cpu/kernels/scale/sve/fp32.cpp')
| -rw-r--r-- | src/cpu/kernels/scale/sve/fp32.cpp | 76 |
1 files changed, 46 insertions, 30 deletions
diff --git a/src/cpu/kernels/scale/sve/fp32.cpp b/src/cpu/kernels/scale/sve/fp32.cpp index f3472f1efd..cbb345edbb 100644 --- a/src/cpu/kernels/scale/sve/fp32.cpp +++ b/src/cpu/kernels/scale/sve/fp32.cpp @@ -25,23 +25,27 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Window.h" + +#include "src/core/helpers/ScaleHelpers.h" #include "src/core/NEON/NEMath.h" #include "src/core/NEON/wrapper/wrapper.h" -#include "src/core/helpers/ScaleHelpers.h" #include "src/core/utils/ScaleUtils.h" #include "support/Rounding.h" +#include <arm_sve.h> #include <cmath> #include <cstddef> -#include <arm_sve.h> - namespace arm_compute { namespace { -void fp32_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets, - float sampling_offset, bool align_corners, const Window &window) +void fp32_sve_scale_nearest(const ITensor *src, + ITensor *dst, + const ITensor *offsets, + float sampling_offset, + bool align_corners, + const Window &window) { const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom; @@ -60,38 +64,50 @@ void fp32_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *off const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3]; - execute_window_loop(win, [&](const Coordinates & id) - { - const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c; - const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr)); - const int offset_row = in_hi * in_stride_wc; - const auto in_ptr = reinterpret_cast<const float *>(in_ptr_start + in_stride_bytes_hwc * id[3]); - const auto out_ptr = reinterpret_cast<float *>(out.ptr()); - - // Compute S elements per iteration - int x = window_start_x; - svbool_t pg = svwhilelt_b32(x, window_end_x); - do + execute_window_loop( + win, + [&](const Coordinates &id) { - // Store results - svst1_f32(pg, out_ptr + x, svld1_f32(pg, in_ptr + offset + offset_row + x)); + const int32_t offset = + *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c; + const auto in_hi = static_cast<int>( + align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) + : std::floor((id.z() + sampling_offset) * hr)); + const int offset_row = in_hi * in_stride_wc; + const auto in_ptr = reinterpret_cast<const float *>(in_ptr_start + in_stride_bytes_hwc * id[3]); + const auto out_ptr = reinterpret_cast<float *>(out.ptr()); - x += svcntw(); - pg = svwhilelt_b32(x, window_end_x); - } - while(svptest_any(svptrue_b32(), pg)); - }, - out); -} + // Compute S elements per iteration + int x = window_start_x; + svbool_t pg = svwhilelt_b32(x, window_end_x); + do + { + // Store results + svst1_f32(pg, out_ptr + x, svld1_f32(pg, in_ptr + offset + offset_row + x)); + + x += svcntw(); + pg = svwhilelt_b32(x, window_end_x); + } while (svptest_any(svptrue_b32(), pg)); + }, + out); } +} // namespace namespace cpu { -void fp32_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, - InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, - bool align_corners, const Window &window) +void fp32_sve_scale(const ITensor *src, + ITensor *dst, + const ITensor *offsets, + const ITensor *dx, + const ITensor *dy, + InterpolationPolicy policy, + BorderMode border_mode, + PixelValue constant_border_value, + float sampling_offset, + bool align_corners, + const Window &window) { ARM_COMPUTE_UNUSED(dx, dy, border_mode, constant_border_value); - if(policy == InterpolationPolicy::NEAREST_NEIGHBOR) + if (policy == InterpolationPolicy::NEAREST_NEIGHBOR) { fp32_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window); } |